Dr. Juan Gómez Luna Prof. Onur Mutlu ETH Zurich Spring 2018

Design of Digital Circuits Lecture 21: SIMD Processors II and Graphics Processing Units
Dr. Juan Gómez Luna Prof. Onur Mutlu ETH Zurich Spring 2018 17 May 2018

New Course: Bachelor’s Seminar in Comp Arch
Fall 2018 2 credit units Rigorous seminar on fundamental and cutting-edge topics in computer architecture Critical presentation, review, and discussion of seminal works in computer architecture We will cover many ideas & issues, analyze their tradeoffs, perform critical thinking and brainstorming Participation, presentation, report and review writing Stay tuned for more information

Agenda for Today & Next Few Lectures
Single-cycle Microarchitectures Multi-cycle and Microprogrammed Microarchitectures Pipelining Issues in Pipelining: Control & Data Dependence Handling, State Maintenance and Recovery, … Out-of-Order Execution Other Execution Paradigms

Readings for Today Peleg and Weiser, “MMX Technology Extension to the Intel Architecture,” IEEE Micro 1996. Lindholm et al., "NVIDIA Tesla: A Unified Graphics and Computing Architecture," IEEE Micro 2008.

Other Approaches to Concurrency (or Instruction Level Parallelism)

Approaches to (Instruction-Level) Concurrency
Pipelining Out-of-order execution Dataflow (at the ISA level) Superscalar Execution VLIW Fine-Grained Multithreading SIMD Processing (Vector and array processors, GPUs) Decoupled Access Execute Systolic Arrays

SIMD Processing: Exploiting Regular (Data) Parallelism

Recall: Flynn’s Taxonomy of Computers
Mike Flynn, “Very High-Speed Computing Systems,” Proc. of IEEE, 1966 SISD: Single instruction operates on single data element SIMD: Single instruction operates on multiple data elements Array processor Vector processor MISD: Multiple instructions operate on single data element Closest form: systolic array processor, streaming processor MIMD: Multiple instructions operate on multiple data elements (multiple instruction streams) Multiprocessor Multithreaded processor

Recall: SIMD Processing
Single instruction operates on multiple data elements In time or in space Multiple processing elements Time-space duality Array processor: Instruction operates on multiple data elements at the same time using different spaces Vector processor: Instruction operates on multiple data elements in consecutive time steps using the same space

Recall: Array vs. Vector Processors
ARRAY PROCESSOR VECTOR PROCESSOR Instruction Stream Same same time Different time LD VR  A[3:0] ADD VR  VR, 1 MUL VR  VR, 2 ST A[3:0]  VR LD0 LD1 LD2 LD3 LD0 AD0 AD1 AD2 AD3 LD1 AD0 MU0 MU1 MU2 MU3 LD2 AD1 MU0 ST0 ST1 ST2 ST3 LD3 AD2 MU1 ST0 AD3 MU2 ST1 Different same space MU3 ST2 Time Same space ST3 Space Space

Recall: Memory Banking
Memory is divided into banks that can be accessed independently; banks share address and data buses (to minimize pin cost) Can start and complete one bank access per cycle Can sustain N parallel accesses if all N go to different banks Bank Bank 1 Bank 2 Bank 15 MDR MAR MDR MAR MDR MAR MDR MAR Data bus Address bus CPU Picture credit: Derek Chiou

Some Issues Stride and banking Storage of a matrix
As long as they are relatively prime to each other and there are enough banks to cover bank access latency, we can sustain 1 element/cycle throughput Storage of a matrix Row major: Consecutive elements in a row are laid out consecutively in memory Column major: Consecutive elements in a column are laid out consecutively in memory You need to change the stride when accessing a row versus column

Matrix Multiplication
A and B, both in row-major order A: Load A0 into vector register V1 Each time, increment address by one to access the next column Accesses have a stride of 1 B: Load B0 into vector register V2 Each time, increment address by 10 Accesses have a stride of 10 A4x6 B6x10 → C4x10 Dot products of rows and columns of A and B Different strides can lead to bank conflicts How do we minimize them?

Minimizing Bank Conflicts
More banks Better data layout to match the access pattern Is this always possible? Better mapping of address to bank E.g., randomized mapping Rau, “Pseudo-randomly interleaved memory,” ISCA 1991.

Recall: Questions (II)
What if vector data is not stored in a strided fashion in memory? (irregular memory access to a vector) Idea: Use indirection to combine/pack elements into vector registers Called scatter/gather operations

Gather/Scatter Operations
Want to vectorize loops with indirect accesses: for (i=0; i<N; i++) A[i] = B[i] + C[D[i]] Indexed load instruction (Gather) LV vD, rD # Load indices in D vector LVI vC, rC, vD # Load indirect from rC base LV vB, rB # Load B vector ADDV.D vA,vB,vC # Do add SV vA, rA # Store result Gather example

Gather/Scatter Operations
Gather/scatter operations often implemented in hardware to handle sparse vectors (matrices) Vector loads and stores use an index vector which is added to the base register to generate the addresses Scatter example Index Vector Data Vector (to Store) Stored Vector (in Memory) Base Base X Base Base X Base X Base X Base Base Scatter example

Array vs. Vector Processors, Revisited
Array vs. vector processor distinction is a “purist’s” distinction Most “modern” SIMD processors are a combination of both They exploit data parallelism in both time and space GPUs are a prime example we will cover in a bit more detail

Recall: Array vs. Vector Processors
ARRAY PROCESSOR VECTOR PROCESSOR Instruction Stream Same same time Different time LD VR  A[3:0] ADD VR  VR, 1 MUL VR  VR, 2 ST A[3:0]  VR LD0 LD1 LD2 LD3 LD0 AD0 AD1 AD2 AD3 LD1 AD0 MU0 MU1 MU2 MU3 LD2 AD1 MU0 ST0 ST1 ST2 ST3 LD3 AD2 MU1 ST0 AD3 MU2 ST1 Different same space MU3 ST2 Time Same space ST3 Space Space

Vector Instruction Execution
VADD A,B  C C[1] C[2] C[0] A[3] B[3] A[4] B[4] A[5] B[5] A[6] B[6] Execution using one pipelined functional unit C[4] C[8] C[0] A[12] B[12] A[16] B[16] A[20] B[20] A[24] B[24] C[5] C[9] C[1] A[13] B[13] A[17] B[17] A[21] B[21] A[25] B[25] C[6] C[10] C[2] A[14] B[14] A[18] B[18] A[22] B[22] A[26] B[26] C[7] C[11] C[3] A[15] B[15] A[19] B[19] A[23] B[23] A[27] B[27] Execution using four pipelined functional units Time Time Space Slide credit: Krste Asanovic

Vector Unit Structure Functional Unit Lane Memory Subsystem
Partitioned Vector Registers Elements 0, 4, 8, … Elements 1, 5, 9, … Elements 2, 6, 10, … Elements 3, 7, 11, … Memory Subsystem Slide credit: Krste Asanovic

Vector Instruction Level Parallelism
Can overlap execution of multiple vector instructions Example machine has 32 elements per vector register and 8 lanes Completes 24 operations/cycle while issuing 1 vector instruction/cycle Load Unit Multiply Unit Add Unit load mul add time load mul add Instruction issue Slide credit: Krste Asanovic

Automatic Code Vectorization
for (i=0; i < N; i++) C[i] = A[i] + B[i]; Vector Instruction load add store Iter. 1 Iter. 2 Vectorized Code Time load add store Iter. 1 Iter. 2 Scalar Sequential Code Vectorization is a compile-time reordering of operation sequencing  requires extensive loop dependence analysis Slide credit: Krste Asanovic

Vector/SIMD Processing Summary
Vector/SIMD machines are good at exploiting regular data-level parallelism Same operation performed on many data elements Improve performance, simplify design (no intra-vector dependencies) Performance improvement limited by vectorizability of code Scalar operations limit vector machine performance Remember Amdahl’s Law CRAY-1 was the fastest SCALAR machine at its time! Many existing ISAs include (vector-like) SIMD operations Intel MMX/SSEn/AVX, PowerPC AltiVec, ARM Advanced SIMD

SIMD Operations in Modern ISAs

SIMD ISA Extensions Single Instruction Multiple Data (SIMD) extension instructions Single instruction acts on multiple pieces of data at once Common application: graphics Perform short arithmetic operations (also called packed arithmetic) For example: add four 8-bit numbers Must modify ALU to eliminate carries between 8-bit values 26

Intel Pentium MMX Operations
Idea: One instruction operates on multiple data elements simultaneously À la array processing (yet much more limited) Designed with multimedia (graphics) operations in mind No VLEN register Opcode determines data type: 8 8-bit bytes 4 16-bit words 2 32-bit doublewords 1 64-bit quadword Stride is always equal to 1. Peleg and Weiser, “MMX Technology Extension to the Intel Architecture,” IEEE Micro, 1996.

MMX Example: Image Overlaying (I)
Goal: Overlay the human in image 1 on top of the background in image 2 Peleg and Weiser, “MMX Technology Extension to the Intel Architecture,” IEEE Micro, 1996.

MMX Example: Image Overlaying (II)
Y = Blossom image X = Woman’s image Peleg and Weiser, “MMX Technology Extension to the Intel Architecture,” IEEE Micro, 1996.

GPUs (Graphics Processing Units)

GPUs are SIMD Engines Underneath
The instruction pipeline operates like a SIMD pipeline (e.g., an array processor) However, the programming is done using threads, NOT SIMD instructions To understand this, let’s go back to our parallelizable code example But, before that, let’s distinguish between Programming Model (Software) vs. Execution Model (Hardware)

Programming Model vs. Hardware Execution Model
Programming Model refers to how the programmer expresses the code E.g., Sequential (von Neumann), Data Parallel (SIMD), Dataflow, Multi-threaded (MIMD, SPMD), … Execution Model refers to how the hardware executes the code underneath E.g., Out-of-order execution, Vector processor, Array processor, Dataflow processor, Multiprocessor, Multithreaded processor, … Execution Model can be very different from the Programming Model E.g., von Neumann model implemented by an OoO processor E.g., SPMD model implemented by a SIMD processor (a GPU)

How Can You Exploit Parallelism Here?
for (i=0; i < N; i++) C[i] = A[i] + B[i]; load add store Iter. 1 Iter. 2 Scalar Sequential Code Let’s examine three programming options to exploit instruction-level parallelism present in this sequential code: 1. Sequential (SISD) 2. Data-Parallel (SIMD) 3. Multithreaded (MIMD/SPMD)

Prog. Model 1: Sequential (SISD)
for (i=0; i < N; i++) C[i] = A[i] + B[i]; Can be executed on a: Pipelined processor Out-of-order execution processor Independent instructions executed when ready Different iterations are present in the instruction window and can execute in parallel in multiple functional units In other words, the loop is dynamically unrolled by the hardware Superscalar or VLIW processor Can fetch and execute multiple instructions per cycle load add store Iter. 1 Iter. 2 Scalar Sequential Code

Prog. Model 2: Data Parallel (SIMD)
for (i=0; i < N; i++) C[i] = A[i] + B[i]; Vectorized Code load add store Iter. 1 Iter. 2 Scalar Sequential Code Vector Instruction load add store load add store VLD A  V1 VLD B  V2 VADD V1 + V2  V3 VST V3  C Iter. 1 Iter. 2 Realization: Each iteration is independent Idea: Programmer or compiler generates a SIMD instruction to execute the same instruction from all iterations across different data Best executed by a SIMD processor (vector, array)

Prog. Model 3: Multithreaded
for (i=0; i < N; i++) C[i] = A[i] + B[i]; load add store Iter. 1 Iter. 2 Scalar Sequential Code load add store load add store Iter. 1 Iter. 2 Realization: Each iteration is independent Idea: Programmer or compiler generates a thread to execute each iteration. Each thread does the same thing (but on different data) Can be executed on a MIMD machine

Prog. Model 3: Multithreaded
for (i=0; i < N; i++) C[i] = A[i] + B[i]; load add store load add store Iter. 1 Iter. 2 Realization: Each iteration is independent Idea: Programmer or compiler generates a thread to execute each iteration. Each thread does the same thing (but on different data) Can be executed on a MIMD machine This particular model is also called: SPMD: Single Program Multiple Data Can be executed on a SIMT machine Single Instruction Multiple Thread Can be executed on a SIMD machine

A GPU is a SIMD (SIMT) Machine
Except it is not programmed using SIMD instructions It is programmed using threads (SPMD programming model) Each thread executes the same code but operates a different piece of data Each thread has its own context (i.e., can be treated/restarted/executed independently) A set of threads executing the same instruction are dynamically grouped into a warp (wavefront) by the hardware A warp is essentially a SIMD operation formed by hardware!

SPMD on SIMT Machine Warp: A set of threads that execute
for (i=0; i < N; i++) C[i] = A[i] + B[i]; load add store load add store Warp 0 at PC X Warp 0 at PC X+1 Warp 0 at PC X+2 Warp 0 at PC X+3 Iter. 1 Iter. 2 Warp: A set of threads that execute the same instruction (i.e., at the same PC) Realization: Each iteration is independent Idea: Programmer or compiler generates a thread to execute each iteration. Each thread does the same thing (but on different data) Can be executed on a MIMD machine This particular model is also called: SPMD: Single Program Multiple Data A GPU executes it using the SIMT model: Single Instruction Multiple Thread Can be executed on a SIMD machine

Graphics Processing Units SIMD not Exposed to Programmer (SIMT)

SIMD vs. SIMT Execution Model
SIMD: A single sequential instruction stream of SIMD instructions  each instruction specifies multiple data inputs [VLD, VLD, VADD, VST], VLEN SIMT: Multiple instruction streams of scalar instructions  threads grouped dynamically into warps [LD, LD, ADD, ST], NumThreads Two Major SIMT Advantages: Can treat each thread separately  i.e., can execute each thread independently (on any type of scalar pipeline)  MIMD processing Can group threads into warps flexibly  i.e., can group threads that are supposed to truly execute the same instruction  dynamically obtain and maximize benefits of SIMD processing

Multithreading of Warps
for (i=0; i < N; i++) C[i] = A[i] + B[i]; Assume a warp consists of 32 threads If you have 32K iterations, and 1 iteration/thread  1K warps Warps can be interleaved on the same pipeline  Fine grained multithreading of warps load add store load add store Warp 1 at PC X Warp 0 at PC X Warp 20 at PC X+2 Iter. 33 Iter. 1 Iter. 20*32 + 1 Iter. 2 Iter. 20*32 + 2 Iter. 34

Warps and Warp-Level FGMT
Warp: A set of threads that execute the same instruction (on different data elements)  SIMT (Nvidia-speak) All threads run the same code Warp: The threads that run lengthwise in a woven fabric … Thread Warp 3 Thread Warp 8 Common PC Thread Warp In SIMD, you need to specify the data array + an instruction (on which to operate the data on) + THE INSTRUCTION WIDTH. Eg: You might want to add 2 integer arrays of length 16, then a SIMD instruction would look like (the instruction has been cooked-up by me for demo) add.16 arr1 arr2 However, SIMT doesn't bother about the instruction width. So, essentially, you could write the above example as: arr1[i] + arr2[i] and then launch as many threads as the length of the array, as you want. Note that, if the array size was, let us say, 32, then SIMD EXPECTS you to explicitly call two such 'add.16' instructions! Whereas, this is not the case with SIMT. Scalar Scalar Scalar Scalar Thread Warp 7 Thread Thread Thread Thread W X Y Z SIMD Pipeline

High-Level View of a GPU

Latency Hiding via Warp-Level FGMT
Warp: A set of threads that execute the same instruction (on different data elements) Fine-grained multithreading One instruction per thread in pipeline at a time (No interlocking) Interleave warp execution to hide latencies Register values of all threads stay in register file FGMT enables long latency tolerance Millions of pixels Decode R F A L U D-Cache Thread Warp 6 Thread Warp 1 Thread Warp 2 Data All Hit? Miss? Warps accessing memory hierarchy Thread Warp 3 Thread Warp 8 Writeback Warps available for scheduling Thread Warp 7 I-Fetch SIMD Pipeline With a large number of shader threads multiplexed on the same execution re- sources, our architecture employs fine-grained multithreading where individual threads are interleaved by the fetch unit to proactively hide the potential latency of stalls before they occur. As illustrated by Figure, warps are issued fairly in a round-robin queue. When a thread is blocked by a memory request, shader core simply removes that thread’s warp from the pool of “ready” warps and thereby allows other threads to proceed while the memory system processes its request. With a large number of threads (1024 per shader core) interleaved on the same pipeline, FGMT effectively hides the latency of most memory operations since the pipeline is occupied with instructions from other threads while memory operations complete. also hides the pipeline latency so that data bypassing logic can potentially be omitted to save area with minimal impact on performance. simplify the dependency check logic design by restricting each thread to have at most one instruction running in the pipeline at any time. Slide credit: Tor Aamodt

Warp Execution (Recall the Slide)
32-thread warp executing ADD A[tid],B[tid]  C[tid] C[1] C[2] C[0] A[3] B[3] A[4] B[4] A[5] B[5] A[6] B[6] Execution using one pipelined functional unit C[4] C[8] C[0] A[12] B[12] A[16] B[16] A[20] B[20] A[24] B[24] C[5] C[9] C[1] A[13] B[13] A[17] B[17] A[21] B[21] A[25] B[25] C[6] C[10] C[2] A[14] B[14] A[18] B[18] A[22] B[22] A[26] B[26] C[7] C[11] C[3] A[15] B[15] A[19] B[19] A[23] B[23] A[27] B[27] Execution using four pipelined functional units Time Time Space Slide credit: Krste Asanovic

SIMD Execution Unit Structure
Functional Unit Lane Registers for each Thread Registers for thread IDs 0, 4, 8, … Registers for thread IDs 1, 5, 9, … Registers for thread IDs 2, 6, 10, … Registers for thread IDs 3, 7, 11, … Memory Subsystem Slide credit: Krste Asanovic

Warp Instruction Level Parallelism
Can overlap execution of multiple instructions Example machine has 32 threads per warp and 8 lanes Completes 24 operations/cycle while issuing 1 warp/cycle Load Unit Multiply Unit Add Unit W0 W1 W2 time W3 W4 W5 Warp issue Slide credit: Krste Asanovic

SIMT Memory Access Same instruction in different threads uses thread id to index and access different data elements Let’s assume N=16, 4 threads per warp  4 warps 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Threads + 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Data elements + + + + Warp 0 Warp 1 Warp 2 Warp 3 Slide credit: Hyesoon Kim

Sample GPU SIMT Code (Simplified)
CPU code for (ii = 0; ii < ; ++ii) { C[ii] = A[ii] + B[ii]; } CUDA code // there are threads __global__ void KernelFunction(…) { int tid = blockDim.x * blockIdx.x + threadIdx.x; int varA = aa[tid]; int varB = bb[tid]; C[tid] = varA + varB; } Slide credit: Hyesoon Kim

Sample GPU Program (Less Simplified)
Slide credit: Hyesoon Kim

Warp-based SIMD vs. Traditional SIMD
Traditional SIMD contains a single thread Sequential instruction execution; lock-step operations in a SIMD instruction Programming model is SIMD (no extra threads)  SW needs to know vector length ISA contains vector/SIMD instructions Warp-based SIMD consists of multiple scalar threads executing in a SIMD manner (i.e., same instruction executed by all threads) Does not have to be lock step Each thread can be treated individually (i.e., placed in a different warp)  programming model not SIMD SW does not need to know vector length Enables multithreading and flexible dynamic grouping of threads ISA is scalar  SIMD operations can be formed dynamically Essentially, it is SPMD programming model implemented on SIMD hardware

SPMD Single procedure/program, multiple data
This is a programming model rather than computer organization Each processing element executes the same procedure, except on different data elements Procedures can synchronize at certain points in program, e.g. barriers Essentially, multiple instruction streams execute the same program Each program/procedure 1) works on different data, 2) can execute a different control-flow path, at run-time Many scientific applications are programmed this way and run on MIMD hardware (multiprocessors) Modern GPUs programmed in a similar way on a SIMD hardware

SIMD vs. SIMT Execution Model
SIMD: A single sequential instruction stream of SIMD instructions  each instruction specifies multiple data inputs [VLD, VLD, VADD, VST], VLEN SIMT: Multiple instruction streams of scalar instructions  threads grouped dynamically into warps [LD, LD, ADD, ST], NumThreads Two Major SIMT Advantages: Can treat each thread separately  i.e., can execute each thread independently on any type of scalar pipeline  MIMD processing Can group threads into warps flexibly  i.e., can group threads that are supposed to truly execute the same instruction  dynamically obtain and maximize benefits of SIMD processing

Threads Can Take Different Paths in Warp-based SIMD
Each thread can have conditional control flow instructions Threads can execute different control flow paths B C D E F A G Thread Warp Common PC Thread 1 Thread 2 Thread 3 Thread 4 Slide credit: Tor Aamodt

Control Flow Problem in GPUs/SIMT
A GPU uses a SIMD pipeline to save area on control logic Groups scalar threads into warps Branch divergence occurs when threads inside warps branch to different execution paths Branch Path A Path B Branch Path A Path B This is the same as conditional/predicated/masked execution. Recall the Vector Mask and Masked Vector Operations? Slide credit: Tor Aamodt

Remember: Each Thread Is Independent
Two Major SIMT Advantages: Can treat each thread separately  i.e., can execute each thread independently on any type of scalar pipeline  MIMD processing Can group threads into warps flexibly  i.e., can group threads that are supposed to truly execute the same instruction  dynamically obtain and maximize benefits of SIMD processing If we have many threads We can find individual threads that are at the same PC And, group them together into a single warp dynamically This reduces “divergence”  improves SIMD utilization SIMD utilization: fraction of SIMD lanes executing a useful operation (i.e., executing an active thread)

Dynamic Warp Formation/Merging
Idea: Dynamically merge threads executing the same instruction (after branch divergence) Form new warps from warps that are waiting Enough threads branching to each path enables the creation of full new warps Warp X Warp Z Warp Y

Dynamic Warp Formation/Merging
Idea: Dynamically merge threads executing the same instruction (after branch divergence) Fung et al., “Dynamic Warp Formation and Scheduling for Efficient GPU Control Flow,” MICRO 2007. Branch Path A Path B Branch Path A

Dynamic Warp Formation Example
Execution of Warp x at Basic Block A Execution of Warp y Legend A B x/1110 y/0011 C x/1000 D x/0110 F x/0001 y/0010 y/0001 y/1100 A new warp created from scalar threads of both Warp x and y executing at Basic Block D D E x/1110 y/0011 G x/1111 y/1111 A A B B C C D D E E F F G G A A Baseline Time Dynamic Warp Formation A A B B C D E E F G G A A Time Slide credit: Tor Aamodt

Hardware Constraints Limit Flexibility of Warp Grouping
Functional Unit Lane Registers for each Thread Registers for thread IDs 0, 4, 8, … Registers for thread IDs 1, 5, 9, … Registers for thread IDs 2, 6, 10, … Registers for thread IDs 3, 7, 11, … Can you move any thread flexibly to any lane? Memory Subsystem Slide credit: Krste Asanovic

Design of Digital Circuits Lecture 21: SIMD Processors II and Graphics Processing Units
Dr. Juan Gómez Luna Prof. Onur Mutlu ETH Zurich Spring 2018 17 May 2018

We did not cover the following slides in lecture
We did not cover the following slides in lecture. These are for your preparation for the next lecture.

An Example GPU

NVIDIA GeForce GTX 285 NVIDIA-speak: 240 stream processors
“SIMT execution” Generic speak: 30 cores 8 SIMD functional units per core Slide credit: Kayvon Fatahalian

NVIDIA GeForce GTX 285 “core”
… 64 KB of storage for thread contexts (registers) 30 * 32 * 32 = 30 * 1024 = 30K fragments 64KB register file = bit registers per thread = 64B (1/32 that of LRB) 16KB of shared scratch 80KB / core available to software = SIMD functional unit, control shared across 8 units = instruction stream decode = multiply-add = execution context storage = multiply Slide credit: Kayvon Fatahalian

NVIDIA GeForce GTX 285 “core”
… 64 KB of storage for thread contexts (registers) To get maximal latency hiding: Run 1/32 of the time 16 words per thread = 64B Groups of 32 threads share instruction stream (each group is a Warp) Up to 32 warps are simultaneously interleaved Up to 1024 thread contexts can be stored Slide credit: Kayvon Fatahalian

30 cores on the GTX 285: 30,720 threads
NVIDIA GeForce GTX 285 Tex Tex … … … … Tex Tex … … Tex Tex … … Tex Tex If you’re running a CUDA program, and your not launching 30K threads, you are certainly not getting full latency hiding, and you might not be using the GPU well … … Tex Tex … … 30 cores on the GTX 285: 30,720 threads Slide credit: Kayvon Fatahalian

Evolution of NVIDIA GPUs

NVIDIA V100 NVIDIA-speak: 5120 stream processors “SIMT execution”
Generic speak: 80 cores 64 SIMD functional units per core Tensor cores for Machine Learning

NVIDIA V100 Block Diagram 80 cores on the V100
80 cores on the V100

NVIDIA V100 Core 15.7 TFLOPS Single Precision 7.8 TFLOPS Double Precision 125 TFLOPS for Deep Learning (Tensor cores)

Dr. Juan Gómez Luna Prof. Onur Mutlu ETH Zurich Spring 2018

Similar presentations

Presentation on theme: "Dr. Juan Gómez Luna Prof. Onur Mutlu ETH Zurich Spring 2018"— Presentation transcript:

Similar presentations

About project

Feedback

Log in

Auth with social network:

Dr. Juan Gómez Luna Prof. Onur Mutlu ETH Zurich Spring 2018

Similar presentations

Presentation on theme: "Dr. Juan Gómez Luna Prof. Onur Mutlu ETH Zurich Spring 2018"— Presentation transcript:

Similar presentations

About project

Feedback